MiniMax released M3.1 Flash, claiming nearly triple the inference speed and 70% lower usage costs than its predecessor.
On October 19, 2026, MiniMax released M3.1 Flash, saying online inference is nearly three times faster than its predecessor and agent-task usage costs are 70% lower. The company lists API rates of $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. It reports 62.8 on Automation-Bench, 73.7 on DeepSWE v1.1, 61.7 on OSWorld 2.0 and 82.2 on CyberGym; output speed can exceed 200 tokens per second, with a median above 170, according to MiniMax. The model has a 1-million-token context window and about 428 billion total parameters, activating 23 billion per token. Its efficiency stack combines MiniMax Sparse Attention, NVFP4 quantization-aware training and DSpark speculative decoding. Lower token prices and faster serving could make extended coding and agent workflows cheaper, intensifying competition among providers of developer models. The launch follows MiniMax M3, released in June with long-context and agentic capabilities, and a run of M2.5 and M2.7 releases earlier in 2026.